Compressive sensing ghost imaging object detection using generative adversarial networks

摘要

Compressive sensing ghost imaging (CSGI) is an imaging mechanism that can nonlocally obtain an unknown object's information with a single-pixel detector by the correlation of intensity fluctuations. In the practical research and application of CSGI, object detection plays a crucial role in real-time monitoring and dynamic optimization of speckle pattern. We demonstrate, for the first time to our knowledge, how to solve the low-resolution and undersampling problems in CSGI object detection. The method we use is to combine generative adversarial networks (GANs) with object detection systems. The robustness of the object detection model can increase by generating reconstructed images of different resolutions and sampling rates for training. The experiment results have verified that the mean average precision of CSGI object detection using GANs has been improved 16.48% and 2.98% on MSCOCO 2017 compared with two traditional learning methods, respectively.